DEV Community: ABINESH. M

Building AI-Powered Data Analytics with Snowflake Cortex

ABINESH. M — Sat, 29 Aug 2026 02:20:04 +0000

Building AI-Powered Data Analytics with Snowflake Cortex

Introduction

Data analytics has evolved far beyond creating dashboards from spreadsheets and traditional databases. Organizations today generate massive amounts of structured and unstructured data, and they need to turn that data into insights quickly.

At the same time, generative AI is changing how people interact with data. Instead of writing complex queries for every question, users increasingly expect to ask questions in natural language, summarize information automatically, and discover patterns that may not be immediately visible.

This is where Snowflake and Snowflake Cortex can work together.

Snowflake provides a modern cloud data platform for storing, processing, and analyzing data, while Cortex brings AI capabilities closer to enterprise data. This combination allows organizations to build analytics workflows where data can be prepared, analyzed, and enhanced with AI within the same platform.

In this article, we'll explore how to approach AI-powered analytics with Snowflake, using a simple sales analytics scenario as an example.

Why Combine Data Analytics and AI?

Traditional analytics generally follows a workflow like this:

Data → SQL → Reports → Human Interpretation

For example, an analyst may have a sales table containing:

Customer information
Product details
Order dates
Revenue
Region
Sales representative

The analyst can use SQL to calculate total revenue, identify the best-performing products, and compare regional performance.

However, there are questions that require additional interpretation.

For example:

"Which products are performing poorly, and what are the likely reasons?"

A SQL query can identify products with declining revenue, but understanding customer feedback or other unstructured information may require AI.

An AI-powered analytics workflow can therefore look like:

Data → SQL Analytics → AI Processing → Insights → Decisions

The important idea is that AI doesn't replace analytics. Instead, AI can complement traditional analytical workflows.

Understanding Snowflake's Role

Snowflake provides the foundation for the analytics workflow.

At a high level, a Snowflake environment can organize data using objects such as:

Database → Schema → Tables → Views

For example:

CREATE DATABASE SALES_ANALYTICS;

CREATE SCHEMA SALES_ANALYTICS.PUBLIC;

A sales table could contain columns such as:

CREATE TABLE SALES_ANALYTICS.PUBLIC.SALES (
    ORDER_ID INTEGER,
    CUSTOMER_NAME VARCHAR,
    PRODUCT_NAME VARCHAR,
    REGION VARCHAR,
    ORDER_DATE DATE,
    QUANTITY INTEGER,
    REVENUE NUMBER(12,2)
);

Once data is available in Snowflake, SQL can be used to perform analytical operations.

For example, to calculate revenue by region:

SELECT
    REGION,
    SUM(REVENUE) AS TOTAL_REVENUE
FROM SALES_ANALYTICS.PUBLIC.SALES
GROUP BY REGION
ORDER BY TOTAL_REVENUE DESC;

This gives analysts a structured view of business performance.

But we can take the workflow further by introducing AI.

Introducing Snowflake Cortex

Snowflake Cortex provides AI and machine learning capabilities that can be used with data stored in Snowflake.

One of the interesting aspects of Cortex is that organizations can bring AI capabilities into their existing data workflows rather than necessarily moving data into a completely separate environment.

Depending on the use case and available features, Cortex can support tasks such as:

Text summarization
Text classification
Sentiment analysis
Natural-language processing
AI-powered applications
Machine learning workflows
Intelligent search and retrieval

This opens up interesting possibilities for analytics teams.

For example, imagine a company stores customer reviews alongside sales data.

Traditional SQL can answer:

"Which products generated the most revenue?"

AI can help answer:

"What are customers saying about those products?"

Combining both creates a much richer analytical workflow.

Step 1: Preparing Analytical Data

Let's imagine our company sells electronic products.

Our sales table contains transactional information:

SELECT
    PRODUCT_NAME,
    REGION,
    SUM(QUANTITY) AS UNITS_SOLD,
    SUM(REVENUE) AS REVENUE
FROM SALES_ANALYTICS.PUBLIC.SALES
GROUP BY PRODUCT_NAME, REGION;

We can use this information to identify high-performing and low-performing products.

For example, we might discover that a particular product has high sales volume in one region but significantly lower performance in another.

This immediately gives the analyst a starting point for investigation.

Step 2: Adding Unstructured Data

Not all valuable business information is structured.

Customer reviews, support tickets, survey responses, and product feedback are examples of unstructured or semi-structured information.

Suppose we have a customer feedback table:

CREATE TABLE CUSTOMER_FEEDBACK (
    FEEDBACK_ID INTEGER,
    PRODUCT_NAME VARCHAR,
    CUSTOMER_REVIEW VARCHAR,
    CREATED_DATE DATE
);

A record might look conceptually like:

Product	Customer Review
Laptop A	Battery life is excellent but the keyboard feels uncomfortable.
Laptop A	Great performance, but the device gets hot during long usage.
Phone B	Camera quality is excellent and the battery lasts all day.

A traditional SQL query can retrieve these reviews, but interpreting hundreds or thousands of reviews manually is inefficient.

This is where AI-powered text analysis becomes useful.

Step 3: Applying AI to Customer Feedback

Snowflake Cortex capabilities can be used to analyze text and generate useful information from unstructured data.

For example, an organization could use AI to determine the sentiment of customer feedback.

Conceptually, the workflow becomes:

Customer Review
↓
Snowflake Table
↓
Cortex AI Analysis
↓
Sentiment / Summary / Classification
↓
Analytics
↓
Business Decision

A feedback dataset could therefore be enriched with AI-generated information such as:

Sentiment
Topic
Summary
Category
Customer intent

Instead of looking at thousands of reviews individually, analysts can aggregate these AI-generated insights.

For example:

SELECT
    PRODUCT_NAME,
    SENTIMENT,
    COUNT(*) AS REVIEW_COUNT
FROM CUSTOMER_FEEDBACK_ANALYZED
GROUP BY PRODUCT_NAME, SENTIMENT
ORDER BY PRODUCT_NAME;

This can help answer questions such as:

Which products receive the most positive feedback?
Which products have the highest negative sentiment?
Which regions generate the most complaints?
What issues are customers mentioning frequently?

Step 4: Combining Structured and Unstructured Data

This is where the workflow becomes especially powerful.

Suppose our sales data shows that Product A generated significant revenue, but customer feedback contains many negative comments about its battery life.

We can combine both datasets.

For example:

SELECT
    S.PRODUCT_NAME,
    SUM(S.REVENUE) AS TOTAL_REVENUE,
    COUNT(F.FEEDBACK_ID) AS TOTAL_FEEDBACK
FROM SALES_ANALYTICS.PUBLIC.SALES S
LEFT JOIN CUSTOMER_FEEDBACK_ANALYZED F
    ON S.PRODUCT_NAME = F.PRODUCT_NAME
GROUP BY S.PRODUCT_NAME
ORDER BY TOTAL_REVENUE DESC;

Now we can compare business performance with customer sentiment.

This creates a more complete picture.

A product may have:

High revenue + positive sentiment

This could indicate a strong product.

But another product might have:

High revenue + negative sentiment

That could indicate an opportunity for product improvement.

Similarly:

Low revenue + positive sentiment

might suggest that the product has potential but requires better marketing or distribution.

The important point is that AI becomes part of the analytical workflow rather than being treated as an isolated chatbot.

Step 5: Creating an AI-Powered Analytics Layer

Once structured and AI-enriched data are available, we can build dashboards and analytical applications on top of them.

A typical architecture could look like:

Data Sources
↓
Sales Data + Customer Reviews + Support Tickets
↓
Snowflake
↓
Data Transformation
↓
SQL Analytics + Cortex AI
↓
Enriched Data
↓
Dashboard + AI Application

The dashboard can provide traditional KPIs such as:

Total revenue
Units sold
Average order value
Regional performance
Product performance

The AI layer can provide additional information:

Customer sentiment
Feedback summaries
Frequently mentioned issues
Text classification
Natural-language insights

This creates a bridge between traditional business intelligence and generative AI.

Why Keeping AI Close to the Data Matters

One of the biggest considerations when building AI applications is data movement.

Enterprise data can contain sensitive business information, customer information, financial information, or internal documents.

Moving data unnecessarily between multiple platforms can introduce additional complexity.

A platform such as Snowflake can provide a centralized environment where data engineering, analytics, governance, and AI workflows can be brought together.

This can simplify architecture and make it easier for teams to work with the same governed data.

For organizations, this means the conversation isn't simply:

"How do we add AI?"

Instead, the more useful question becomes:

"How do we integrate AI into the data workflows we already trust?"

Practical Use Cases

1. Retail

Retail organizations can combine transaction data with customer reviews.

AI can help identify:

Product complaints
Customer sentiment
Frequently requested features
Emerging trends

Analysts can then compare those insights with sales performance.

2. Financial Services

Financial organizations can analyze large volumes of documents and text while combining them with structured financial data.

Potential applications include:

Document summarization
Classification
Customer communication analysis
Risk-related text analysis

3. Healthcare

Healthcare organizations deal with both structured and unstructured information.

AI-powered text analysis can potentially help summarize or classify large volumes of appropriate textual data while analytics systems provide structured reporting.

4. Customer Support

Support teams can analyze tickets automatically.

Instead of simply counting tickets, AI can help classify them into categories such as:

Billing
Technical issue
Product request
Account problem

Analytics teams can then track these categories over time.

Best Practices for AI-Powered Analytics

Building an AI-powered analytics solution isn't just about calling an AI function.

A successful implementation should consider several areas.

1. Start with a clear business problem

Don't add AI simply because AI is popular.

Identify a problem where AI provides measurable value.

2. Prepare your data carefully

Poor-quality data can produce poor analytical results.

Data validation, transformation, and governance remain important even when AI is involved.

3. Validate AI-generated results

AI output should not automatically be treated as fact.

For important business decisions, organizations should implement appropriate validation and human review.

4. Protect sensitive information

Understand what data is being processed and apply appropriate access controls and governance policies.

5. Monitor cost and performance

AI workloads can introduce additional computational requirements.

Teams should monitor usage and optimize workflows where appropriate.

6. Combine AI with traditional analytics

AI should complement SQL, dashboards, statistical analysis, and business intelligence rather than replacing them completely.

From Dashboard to Intelligent Data Platform

Traditional analytics answers questions such as:

"What happened?"

Advanced analytics can help answer:

"Why did it happen?"

AI-powered analytics can go one step further:

"What does the available information suggest, and what should we investigate next?"

This progression demonstrates why combining analytics and AI is becoming increasingly important.

Snowflake provides the data foundation, while Cortex capabilities can help bring AI into workflows involving structured and unstructured information.

Instead of maintaining completely disconnected systems for data storage, analytics, and AI experimentation, teams can build more integrated workflows around their governed data.

Conclusion

AI-powered analytics is not about replacing data analysts with AI. It is about giving analysts and organizations better tools to work with increasingly complex data.

Snowflake provides a strong foundation for storing, transforming, and analyzing data. Snowflake Cortex extends that environment with AI capabilities that can help organizations work with text, generate insights, and build intelligent data applications.

The most interesting opportunities appear when these capabilities are combined.

A simple sales dashboard might tell us that a product is underperforming. Customer feedback analyzed with AI might help us understand why. Bringing both pieces of information together gives decision-makers a much more complete picture.

For data professionals, this creates an exciting direction:

SQL + Data Engineering + Analytics + AI

Learning how these areas work together can help us move from simply analyzing data to building intelligent data solutions.

The future of analytics isn't just about having more data.

It's about making that data more useful.

Explore Generative AI with the Gemini API in Vertex AI

ABINESH. M — Sat, 19 Jul 2025 07:44:28 +0000

🤖 Explore Generative AI with the Gemini API in Vertex AI
The future of intelligent applications is being shaped by generative AI. With Google Cloud’s Vertex AI and its flagship Gemini API, developers now have access to powerful multimodal models capable of understanding and generating text, images, code, and more.

In this blog, we’ll explore:

What Gemini is and why it matters

How to access and use the Gemini API via Vertex AI

Example use cases (with code!)

Best practices for performance and safety

How to start building your own GenAI apps

🌟 What is Gemini?
Gemini is Google DeepMind’s family of multimodal large language models (LLMs), designed to understand and generate across:

📝 Natural language

💻 Programming code

🖼️ Images (Gemini 1.5 Pro and later)

📄 Documents (PDFs, slides, etc.)

The Gemini API, integrated with Vertex AI, allows developers to use these models via Python, REST, or in Vertex AI Studio—a no-code playground for testing prompts.

⚙️ Why Vertex AI?
Vertex AI is Google Cloud’s unified ML platform. It lets you:

Access foundation models like Gemini via API

Tune models with adapters or prompt engineering

Integrate LLMs with your apps, pipelines, and workflows

Monitor usage, safety, and cost with enterprise-grade tooling

Gemini models on Vertex AI support text-only and multimodal inputs, depending on the variant (e.g., Gemini 1.5 Pro supports up to 1M tokens and image input).

🚀 Getting Started with Gemini API
✅ Step 1: Enable Vertex AI API
Go to the Google Cloud Console

Enable Vertex AI API and Generative AI support

✅ Step 2: Install Python SDK
bash
Copy
Edit
pip install google-cloud-aiplatform
✅ Step 3: Authenticate and Initialize
python
Copy
Edit
from vertexai.preview.generative_models import GenerativeModel
import vertexai

vertexai.init(project="your-gcp-project-id", location="us-central1")
💡 Example: Ask Gemini to Summarize
python
Copy
Edit
model = GenerativeModel("gemini-1.5-pro")

response = model.generate_content("Summarize the key points of the Paris Climate Agreement.")
print(response.text)
✅ Gemini responds with a clear, multi-paragraph summary.

🧠 Advanced: Multimodal Input Example
Gemini 1.5 Pro supports image + text prompts.

python
Copy
Edit
with open("chart.png", "rb") as image_file:
response = model.generate_content(
[
"What trend is shown in this chart?",
],
files={"image": image_file}
)

print(response.text)
Use cases:

Visual document Q&A

UI/UX screenshot analysis

Marketing asset feedback

🧰 Use Cases in the Real World
Industry GenAI Task with Gemini
🏥 Healthcare Summarize patient records (text + chart)
🏛️ Legal Analyze contracts and flag clauses
📊 Finance Visualize trends in reports
📚 EdTech Tutor bots that generate and explain
🛍️ E-commerce Auto-generate product descriptions
🤖 DevTools Explain, refactor, or write code

🛡️ Best Practices for Using Gemini API
🔐 Safety first: Use safety filters and review output policies

⚙️ Tune settings: Experiment with temperature, top-k, and max tokens

🧪 Prompt iterate: Refine prompts for clarity and accuracy

📦 Chunk large content: For long docs, split into meaningful sections

📈 Monitor performance: Use Vertex AI metrics dashboard

💬 Pro Tip: Use Gemini in Vertex AI Studio
Want a low-code way to test Gemini?

Go to Vertex AI Studio

Select Gemini 1.5 Pro

Start prompting immediately with text, files, or images

Great for prototyping before production deployment.

🔚 Conclusion
The Gemini API in Vertex AI gives you access to one of the most advanced LLMs available—directly in your app stack. Whether you’re building an AI chatbot, summarizing legal documents, or generating social media copy, Gemini can handle the logic, language, and visuals behind it all.

With just a few lines of code, you're no longer just using AI—you're building with it.

Inspect Rich Documents with Gemini Multimodality and Multimodal RAG

ABINESH. M — Sat, 19 Jul 2025 07:43:18 +0000

📄 Inspect Rich Documents with Gemini Multimodality and Multimodal RAG
As enterprise data becomes increasingly complex, the need to analyze rich documents—such as PDFs, images, tables, scanned forms, and reports—has never been more urgent. Traditional text-based models fall short when faced with visual or structured content. That’s where Gemini’s multimodal capabilities and Multimodal RAG (Retrieval-Augmented Generation) come in.

In this article, you'll learn:

What Gemini multimodality offers

Why traditional RAG struggles with rich content

How Multimodal RAG solves this problem

Real-world use cases

How to implement a basic inspection pipeline using Gemini 1.5 Pro

🌐 Gemini Multimodality: More Than Just Text
Google's Gemini 1.5 Pro, available in Vertex AI, is a multimodal large language model (MLLM) that can accept combinations of:

🧾 Text

🖼️ Images

📄 PDFs

📊 Tables

📁 Code snippets

It can:

Read and interpret scanned documents

Understand visual layouts and complex tables

Cross-reference data across images and text

Analyze charts and structured forms

This makes it ideal for document intelligence tasks—especially when those documents go beyond plain text.

🔍 What Is Multimodal RAG?
Retrieval-Augmented Generation (RAG) improves LLM accuracy by retrieving relevant documents or content from a database before passing it to the model. Multimodal RAG takes this a step further by:

Indexing and retrieving images, PDFs, tables, or a mix of modalities

Letting the model reason over text and visuals together

Enabling context-aware QA from complex data

📘 Example: Given a 20-page financial report PDF with charts and footnotes, Multimodal RAG enables Gemini to:

Retrieve relevant sections and visuals

Understand the data points from charts

Answer “What is the net profit trend over the last 3 years?”

🧠 Real-World Use Cases
Industry Use Case
🏥 Healthcare Extract insights from medical forms and x-rays
💼 Legal Summarize and compare legal contracts
📊 Finance Analyze quarterly reports and charts
🏗️ Manufacturing Understand scanned checklists and invoices
🏛️ Government Process handwritten forms and old records

🛠️ How to Implement Gemini + Multimodal RAG
Here’s how you can build a simple Multimodal RAG pipeline using Gemini:

Preprocess & Chunk Documents Use pdfplumber, PyMuPDF, or Unstructured.io to extract text & images from PDFs

Store structured chunks in a vector DB like FAISS, Weaviate, or Pinecone

python
Copy
Edit
from unstructured.partition.pdf import partition_pdf
chunks = partition_pdf("report.pdf") # returns text + image segments

Embed & Store in Vector DB
Use multimodal embeddings or store image paths and chunk metadata.
Retrieve Relevant Chunks
When a query is entered, retrieve relevant document snippets (text or image-based).

python
Copy
Edit
query = "What is the revenue growth from 2020 to 2023?"
results = vector_db.search(query, top_k=5)

Pass to Gemini 1.5 Pro with Context Gemini supports file input via Vertex AI SDK:

python
Copy
Edit
from vertexai.generative_models import GenerativeModel

model = GenerativeModel("gemini-1.5-pro")

response = model.generate_content(
[
"Answer this question based on the uploaded document:",
f"Question: {query}"
],
files={"document": open("chunk1.pdf", "rb")}
)
print(response.text)
You can pass multiple files (images, CSVs, etc.) together.

💡 Best Practices for Rich Document QA
🧠 Add OCR for scanned files (e.g., Tesseract or Google Document AI)

🧩 Use chunk overlap to preserve context

🧾 Maintain layout by storing positional metadata (X-Y axis from PDFs)

📦 Compress large PDFs or resize images before sending to Gemini

🚀 Power Use Case: Board Meeting Intelligence Tool
Imagine uploading:

30-page PDF board meeting slides

A ZIP file of Excel budget sheets

Product screenshots (JPG)

A Word doc of notes

And asking:

“Summarize our revenue performance, budget allocation changes, and product roadmap updates.”

Multimodal RAG with Gemini can piece all of that together—text, images, and tables—and give you one cohesive answer.

🔚 Conclusion
Inspecting rich documents isn’t just about reading text. It’s about interpreting layout, visuals, structure, and relationships across modalities. With Gemini's multimodal capabilities and a Multimodal RAG approach, you can build intelligent document processing pipelines for almost any industry.

Start today with Gemini in Vertex AI Studio, or build your own app with the Python SDK.